I started programming my CFD simulations in parallel to use with a Linux Cluster. I can run Serial simulations just fine, and I can run Parallel 2 CPU simulations fine as well. When I try to increase the number of CPUs higher than 2, the simulation halts when the iterations begin.

At first I thought it might be some of the parallel coding that I added, but I went back to the original serial version and I get the exact same result. It makes no difference if I spawn 4 nodes on 1 dual processor machine or spawn 1 node per CPU on 2 different dual processor machines (thought it might have been a communication issue).

Is there anything obvious that I should be looking for to fix in my UDF file? I am launching Fluent by:

fluent 2d -g -t4 -alnamd64

The -alnamd64 switch is needed for that particular cluster. I tried the switch -psmpi as well, but that didn't help.

Any suggestions?

try -pnmpi. you will need to use -cnf=hostname file path with the -pnmpi switch. hostname file is just text file with host names eg. node1 node2 etc

try partitioning the grid in serial before running in parallel

Thanks for the responses. I tried the -pnmpi switch and I also tried partitioning the grid in serial before running in parallel, but neither seemed to help

Is there anything else (that I may have written in my UDF) obvious which would cause the solver to get stuck? I thought about File I/O, but I disabled any file read/write and that didn't help.

The UDF basically records velocities from particular cells (C_U,V,W), computes an output, and applies that value as a boundary input (F_PROFILE). I suppose I could try (as a test) getting rid of the UDF completely and see if Fluent still gets stuck....

Any more suggestions?

hi try to spawn manually your nodes, and don't type -t4 option. It should be done before reading your case

How do I manually spawn nodes from Fluent without the GUI? I've read Ch. 30 (Parallel Processing) of the Fluent Manual, but it only mentions spawning nodes through the GUI.

I created a test simulation which did not use any UDFs and I get the exact same results when going from 2 to 4 CPU's (simulation "locks up" upon beginning iterations - as if it is waiting for something....), so at least that narrows down the possibilities.

It is most likely a communication issue on the cluster itself, but I'm not sure how to get around it.....

try with gui. you will know immediatly if your cluster is well set up. If it's ok, then write a journal while spawning your nodes. After that you will be able to spwan your nodes "manually" and automatically. It's like a -cnf options with hostfile